Towards Content Transfer through Grounded Text Generation

This repo contains the code and data of the following paper:

Towards Content Transfer through Grounded Text Generation. Shrimai Prabhumoye, Chris Quirk, Michel Galley. NAACL 2019. arXiv

Dependencies

• Python 3.6
• Pytorch 0.3
• sentencepiece
• NLTK
• nltk.download("stopwords") in your python terminal.

Data

Dowload the train, dev, and test data for all the experiments from the following link:

http://tts.speech.cs.cmu.edu/content_transfer/train_data.zip
unzip train_data.zip

The *.src files contain the news articles, *.cxt files contain the Wikipedia context, *.tgt files contain the target sentences and the *.srcxt files contain the news articles concatenated with Wikipedia context used in CAG models.

Dowload the raw data for train, dev, and test splits from the following link:

http://tts.speech.cs.cmu.edu/content_transfer/raw_data.zip
unzip raw_data.zip

The raw data gives the following information:

1. wikiID: Wikipedia page ID
2. wikiTitle: Wikipedia page Title
3. wikiContext: Context of the Wikipedia article as is. This is a list of list of sentences.
4. Target: Target sentence from the Wikipedia article as is.
5. clean_wikiContext: The cleaned version of the Wikipedia context. This is a list of sentences.
6. clean_Target: The cleaned version of the target sentence.
7. domain: The domain of the news article.
8. URL: The URL of the news article.
9. curlCommand: The curl command to download the news article from common crawl.
10. HTML_Text: HTML of the news article converted to plain text.
11. clean_HTML_Text: Clean version of the plain text.

The domains.txt file contains the list of domains used to collect the dataset.

Models

• Trained sentencepiece model

Download the trained sentencepeice model used in all experiments.

http://tts.speech.cs.cmu.edu/content_transfer/sentencepieceModel.zip
unzip sentencepieceModel.zip

To train sentencepiece model on your data:

python sentence_piece.py -mode train -input sentencepieceModel/train.data -model_prefix testModel -model_type bpe -vocab_size 32000

To encode data using the trained sentencepiece model:

python sentence_piece.py -mode encode -input inputFilename.txt -model sentencepieceModel/bpeM.model -output outputFilename.txt

To decode the generated data using the trained sentencepiece model:

python sentence_piece.py -mode decode -input inputFilename.txt -output outputFilename.txt -model sentencepieceModel/bpeM.model

• Sum-Basic(SB) and Context Informed Sum-Basic (CISB)

python sumbasicUpdate.py -input raw_data/filename.csv -output filename.txt

Use the -context_update flag for CISB.

• Context Agnostic Generative (CAG) Model and Context Informed Generative (CIG) Model

Please use the code base in the following git repo for these two models: https://github.com/shrimai/Style-Transfer-Through-Back-Translation Refer to example.sh file to see the commands.

Download the trained CAG model:

http://tts.speech.cs.cmu.edu/content_transfer/cag_model.zip
unzip cag_model.zip

Download the trained CIG model:

http://tts.speech.cs.cmu.edu/content_transfer/cig_model.zip
unzip cig_model.zip

• Context Receptive Generative (CRG) Model

Follow the example.sh file in the context_receptive_generative/ directory.

Download the trained CRG model:

http://tts.speech.cs.cmu.edu/content_transfer/crg_model.zip
unzip crg_model.zip

If you are using this data or code then please cite the following paper::

@inproceedings{content_transfer_naacl19,
title={Towards Content Transfer through Grounded Text Generation},
author={Prabhumoye, Shrimai and Quirk, Chris and Galley, Michel},
year={2019},
booktitle={Proc. NAACL}
}